THE IXM2 PARALLEL ASSOCIATIVE PROCESSOR FOR AI

Although researchers have been studying associative memory and associa tive processing for 30 years, the importance of associative memories a s SIMD (single instruction, multiple data) devices has not been suffic iently recognized. Part of this is because large associative memories are difficult to develop, and the small associatiative memories availa ble so far have allowed much less parallelism than needed to demonstra te effective SIMD processing on associative memories. The authors desc ribe the IXM2 associative processor and its main application, speech-t o-speech translation. The IXM2 began as a faithful implementation of t he NETL semantic network machine and grew into a massively parallel SI MD machine that demonstrates the power of large associative memories. The social implications of successful automatic translation are enormo us. Persons who speak different languages could communicate in real ti me by using interpreting telephony. This translation involves many com puting operations because it requires such features as real-time respo nse and a very large capacity for handling vocabulary. The article foc uses on memory-based models of speech-to-speech translation that view episodic memory as the foundation of intelligence. This view naturally leads to implementing intelligent associative memories. The authors f irst discuss the application as an illustration of associative process ing and then shift to an exploration of the IXM2 itself, giving perfor mance results.